Machine Learning Study Group: Papers and Courses Together

A machine learning study group is a few people who either take one course on a shared schedule or read one research paper a week and discuss it. Pick a course while most of you are learning the basics, and papers once everyone can follow a model's training code. Open reading groups and free course cohorts both meet online every week.

A course or papers?

Whichever you choose, do the math alone and check it together. Mathematics for Machine Learning, by Marc Peter Deisenroth, A. Aldo Faisal and Cheng Soon Ong, is free as a PDF. Its first seven chapters cover linear algebra, analytic geometry, matrix decompositions, vector calculus, probability and optimization. Work a chapter in a quiet session during the week and bring only the exercises nobody could solve.

Reading groups and cohorts running now

These were running when we checked on November 14, 2026:

To join DLCT, sign up for its Google Group, which emails the Zoom link each week. The Latent Space club grew out of the Latent Space Discord, which had 14,527 members that day. On StudyDuo, Find Partners includes Machine learning and Artificial intelligence among its interests, so you can search for people by either one. In any Discord or Slack group, keep your phone number and logins private.

How a machine learning study group reads a paper in an hour

S. Keshav's three-pass method splits reading a paper into stages (ACM SIGCOMM Computer Communication Review, 2007). The first pass covers the title, abstract, introduction, headings and conclusions. It ends with five questions: what kind of paper it is, what it builds on, whether its assumptions look valid, what it contributes, and whether it's clear. The second pass adds the figures, checking axes and error bars, and marks the references you haven't read. In the third, you try to re-create the work yourself, and that's the expensive one:

So split the work. Everyone does the first two passes before the meeting, about an hour and ten minutes, and only that week's presenter does the third. Keshav's test for the second pass is whether you can summarize the paper's main thrust, with its evidence, to someone else, and the meeting's opening round checks exactly that. The Latent Space club works on the same principle: "Together, we pre-read and discuss a paper weekly."

The others can each take a role. Alec Jacobson and Colin Raffel describe role-playing paper-reading seminars, a format that grew from one Jacobson designed with Eitan Grinspun at Columbia. Each member plays one of seven roles for the week, and the roles rotate. A group of four or five can use these:

• Peer reviewer: writes a short review and recommends accepting or rejecting the paper.
• Archaeologist: finds one older paper this one builds on and one newer paper that cites it.
• Hacker: codes a stripped-down version of the method on a toy problem and shares the core of it. Two people can pair program this part.
• Industry practitioner: argues for or against building the method into a product.

What keeps a reading group going

Medicine has studied reading groups for decades under the name journal club, and the findings carry over to a machine learning study group. Yogeswari Deenadayalan and colleagues reviewed 101 articles on them, 21 of which formed the body of evidence (Journal of Evaluation in Clinical Practice, 2008). Over 80% of the papers reported gains in knowledge and critical appraisal skills. Successful clubs shared regular, expected meetings, a clear purpose, a leader who picked papers and ran the discussion, papers sent out in advance, one appraisal routine and written summaries.

An older survey asked what makes a club last. Jaan Sidorov surveyed the chief residents of 208 internal medicine programs in the eastern US in 1992 and heard back from 131 (Archives of Internal Medicine, 1995). Two-thirds of their clubs, 65.6%, had run for two years or more, and attendance averaged 58.5%. Smaller programs and mandatory attendance went with higher turnout, and regular food went with a longer life. It was a survey of medical residents, so it shows associations, not causes.

Online, that evidence translates into five habits:

1. Keep one slot every week, even when only two people can come.
2. Send the paper seven days ahead, chosen by that week's lead.
3. Use the same routine each time: the five first-pass questions, then the roles.
4. Keep a running summary, three lines per paper: the claim, the evidence, your doubt.
5. Stay small and committed: four to six people who either come or hand their role to someone else. In person, bring food.

Between sessions

In a machine learning study group, the third pass, the math chapters and the course exercises are solo work. StudyDuo, which we make, is built for that kind of hour. Focus Room never closes: name the paper or the lesson in your task and set 30 to 120 minutes. Then read on camera beside others doing their own work, with every mic muted. Study Hour, once a day, gives your group a single focused hour with no breaks.

Before presenting, rehearse the walkthrough out loud in a 1:1, where both of you can talk; 1:1s need a verified photo. Explaining a method to a listener is the Feynman technique applied to a paper. Groups that want a room for every member can start from the online study group page. If yours would rather build SQL, statistics and a portfolio, read Data Science Study Group.